Benchmarking LLMs' Swarm intelligence

Large Language Models (LLMs) show potential for complex reasoning, yet their capacity for emergent coordination in Multi-Agent Systems (MAS) when operating under strict constraints-such as limited local perception and communication, characteristic of natural swarms-remains largely unexplored, particularly concerning the nuances of swarm intelligence. Existing benchmarks often do not fully capture the unique challenges of decentralized coordination that arise when agents operate with incomplete spatio-temporal information. To bridge this gap, we introduce SwarmBench, a novel benchmark designed to systematically evaluate the swarm intelligence capabilities of LLMs acting as decentralized agents. SwarmBench features five foundational MAS coordination tasks within a configurable 2D grid environment, forcing agents to rely primarily on local sensory input (k x k view) and local communication. We propose metrics for coordination effectiveness and analyze emergent group dynamics. Evaluating several leading LLMs in a zero-shot setting, we find significant performance variations across tasks, highlighting the difficulties posed by local information constraints. While some coordination emerges, results indicate limitations in robust planning and strategy formation under uncertainty in these decentralized scenarios. Assessing LLMs under swarm-like conditions is crucial for realizing their potential in future decentralized systems. We release SwarmBench as an open, extensible toolkit-built upon a customizable and scalable physical system with defined mechanical properties. It provides environments, prompts, evaluation scripts, and the comprehensive experimental datasets generated, aiming to foster reproducible research into LLM-based MAS coordination and the theoretical underpinnings of Embodied MAS. Our code repository is available at https://github.com/x66ccff/swarmbench.

DeepCritic: Deliberate Critique with Large Language Models

As Large Language Models (LLMs) are rapidly evolving, providing accurate feedback and scalable oversight on their outputs becomes an urgent and critical problem. Leveraging LLMs as critique models to achieve automated supervision is a promising solution. In this work, we focus on studying and enhancing the math critique ability of LLMs. Current LLM critics provide critiques that are too shallow and superficial on each step, leading to low judgment accuracy and struggling to offer sufficient feedback for the LLM generator to correct mistakes. To tackle this issue, we propose a novel and effective two-stage framework to develop LLM critics that are capable of deliberately critiquing on each reasoning step of math solutions. In the first stage, we utilize Qwen2.5-72B-Instruct to generate 4.5K long-form critiques as seed data for supervised fine-tuning. Each seed critique consists of deliberate step-wise critiques that includes multi-perspective verifications as well as in-depth critiques of initial critiques for each reasoning step. Then, we perform reinforcement learning on the fine-tuned model with either existing human-labeled data from PRM800K or our automatically annotated data obtained via Monte Carlo sampling-based correctness estimation, to further incentivize its critique ability. Our developed critique model built on Qwen2.5-7B-Instruct not only significantly outperforms existing LLM critics (including the same-sized DeepSeek-R1-distill models and GPT-4o) on various error identification benchmarks, but also more effectively helps the LLM generator refine erroneous steps through more detailed feedback.

WebThinker: Empowering Large Reasoning Models with Deep Research Capability

Large reasoning models (LRMs), such as OpenAI-o1 and DeepSeek-R1, demonstrate impressive long-horizon reasoning capabilities. However, their reliance on static internal knowledge limits their performance on complex, knowledge-intensive tasks and hinders their ability to produce comprehensive research reports requiring synthesis of diverse web information. To address this, we propose \textbf{WebThinker}, a deep research agent that empowers LRMs to autonomously search the web, navigate web pages, and draft research reports during the reasoning process. WebThinker integrates a \textbf{Deep Web Explorer} module, enabling LRMs to dynamically search, navigate, and extract information from the web when encountering knowledge gaps. It also employs an \textbf{Autonomous Think-Search-and-Draft strategy}, allowing the model to seamlessly interleave reasoning, information gathering, and report writing in real time. To further enhance research tool utilization, we introduce an \textbf{RL-based training strategy} via iterative online Direct Preference Optimization (DPO). Extensive experiments on complex reasoning benchmarks (GPQA, GAIA, WebWalkerQA, HLE) and scientific report generation tasks (Glaive) demonstrate that WebThinker significantly outperforms existing methods and strong proprietary systems. Our approach enhances LRM reliability and applicability in complex scenarios, paving the way for more capable and versatile deep research systems. The code is available at https://github.com/RUC-NLPIR/WebThinker.

Search-Based Interaction For Conversation Recommendation via Generative Reward Model Based Simulated User

Conversational recommendation systems (CRSs) use multi-turn interaction to capture user preferences and provide personalized recommendations. A fundamental challenge in CRSs lies in effectively understanding user preferences from conversations. User preferences can be multifaceted and complex, posing significant challenges for accurate recommendations even with access to abundant external knowledge. While interaction with users can clarify their true preferences, frequent user involvement can lead to a degraded user experience. To address this problem, we propose a generative reward model based simulated user, named GRSU, for automatic interaction with CRSs. The simulated user provides feedback to the items recommended by CRSs, enabling them to better capture intricate user preferences through multi-turn interaction. Inspired by generative reward models, we design two types of feedback actions for the simulated user: i.e., generative item scoring, which offers coarse-grained feedback, and attribute-based item critique, which provides fine-grained feedback. To ensure seamless integration, these feedback actions are unified into an instruction-based format, allowing the development of a unified simulated user via instruction tuning on synthesized data. With this simulated user, automatic multi-turn interaction with CRSs can be effectively conducted. Furthermore, to strike a balance between effectiveness and efficiency, we draw inspiration from the paradigm of reward-guided search in complex reasoning tasks and employ beam search for the interaction process. On top of this, we propose an efficient candidate ranking method to improve the recommendation results derived from interaction. Extensive experiments on public datasets demonstrate the effectiveness, efficiency, and transferability of our approach.

Slow Thinking for Sequential Recommendation

To develop effective sequential recommender systems, numerous methods have been proposed to model historical user behaviors. Despite the effectiveness, these methods share the same fast thinking paradigm. That is, for making recommendations, these methods typically encodes user historical interactions to obtain user representations and directly match these representations with candidate item representations. However, due to the limited capacity of traditional lightweight recommendation models, this one-step inference paradigm often leads to suboptimal performance. To tackle this issue, we present a novel slow thinking recommendation model, named STREAM-Rec. Our approach is capable of analyzing historical user behavior, generating a multi-step, deliberative reasoning process, and ultimately delivering personalized recommendations. In particular, we focus on two key challenges: (1) identifying the suitable reasoning patterns in recommender systems, and (2) exploring how to effectively stimulate the reasoning capabilities of traditional recommenders. To this end, we introduce a three-stage training framework. In the first stage, the model is pretrained on large-scale user behavior data to learn behavior patterns and capture long-range dependencies. In the second stage, we design an iterative inference algorithm to annotate suitable reasoning traces by progressively refining the model predictions. This annotated data is then used to fine-tune the model. Finally, in the third stage, we apply reinforcement learning to further enhance the model generalization ability. Extensive experiments validate the effectiveness of our proposed method.

How do Large Language Models Understand Relevance? A Mechanistic Interpretability Perspective

Recent studies have shown that large language models (LLMs) can assess relevance and support information retrieval (IR) tasks such as document ranking and relevance judgment generation. However, the internal mechanisms by which off-the-shelf LLMs understand and operationalize relevance remain largely unexplored. In this paper, we systematically investigate how different LLM modules contribute to relevance judgment through the lens of mechanistic interpretability. Using activation patching techniques, we analyze the roles of various model components and identify a multi-stage, progressive process in generating either pointwise or pairwise relevance judgment. Specifically, LLMs first extract query and document information in the early layers, then process relevance information according to instructions in the middle layers, and finally utilize specific attention heads in the later layers to generate relevance judgments in the required format. Our findings provide insights into the mechanisms underlying relevance assessment in LLMs, offering valuable implications for future research on leveraging LLMs for IR tasks.

Pre-training Generative Recommender with Multi-Identifier Item Tokenization

Generative recommendation autoregressively generates item identifiers to recommend potential items. Existing methods typically adopt a one-to-one mapping strategy, where each item is represented by a single identifier. However, this scheme poses issues, such as suboptimal semantic modeling for low-frequency items and limited diversity in token sequence data. To overcome these limitations, we propose MTGRec, which leverages Multi-identifier item Tokenization to augment token sequence data for Generative Recommender pre-training. Our approach involves two key innovations: multi-identifier item tokenization and curriculum recommender pre-training. For multi-identifier item tokenization, we leverage the RQ-VAE as the tokenizer backbone and treat model checkpoints from adjacent training epochs as semantically relevant tokenizers. This allows each item to be associated with multiple identifiers, enabling a single user interaction sequence to be converted into several token sequences as different data groups. For curriculum recommender pre-training, we introduce a curriculum learning scheme guided by data influence estimation, dynamically adjusting the sampling probability of each data group during recommender pre-training. After pre-training, we fine-tune the model using a single tokenizer to ensure accurate item identification for recommendation. Extensive experiments on three public benchmark datasets demonstrate that MTGRec significantly outperforms both traditional and generative recommendation baselines in terms of effectiveness and scalability.

Universal Item Tokenization for Transferable Generative Recommendation

Recently, generative recommendation has emerged as a promising paradigm, attracting significant research attention. The basic framework involves an item tokenizer, which represents each item as a sequence of codes serving as its identifier, and a generative recommender that predicts the next item by autoregressively generating the target item identifier. However, in existing methods, both the tokenizer and the recommender are typically domain-specific, limiting their ability for effective transfer or adaptation to new domains. To this end, we propose UTGRec, a Universal item Tokenization approach for transferable Generative Recommendation. Specifically, we design a universal item tokenizer for encoding rich item semantics by adapting a multimodal large language model (MLLM). By devising tree-structured codebooks, we discretize content representations into corresponding codes for item tokenization. To effectively learn the universal item tokenizer on multiple domains, we introduce two key techniques in our approach. For raw content reconstruction, we employ dual lightweight decoders to reconstruct item text and images from discrete representations to capture general knowledge embedded in the content. For collaborative knowledge integration, we assume that co-occurring items are similar and integrate collaborative signals through co-occurrence alignment and reconstruction. Finally, we present a joint learning framework to pre-train and adapt the transferable generative recommender across multiple domains. Extensive experiments on four public datasets demonstrate the superiority of UTGRec compared to both traditional and generative recommendation baselines.

Test-Time Alignment for Tracking User Interest Shifts in Sequential Recommendation

Sequential recommendation is essential in modern recommender systems, aiming to predict the next item a user may interact with based on their historical behaviors. However, real-world scenarios are often dynamic and subject to shifts in user interests. Conventional sequential recommendation models are typically trained on static historical data, limiting their ability to adapt to such shifts and resulting in significant performance degradation during testing. Recently, Test-Time Training (TTT) has emerged as a promising paradigm, enabling pre-trained models to dynamically adapt to test data by leveraging unlabeled examples during testing. However, applying TTT to effectively track and address user interest shifts in recommender systems remains an open and challenging problem. Key challenges include how to capture temporal information effectively and explicitly identifying shifts in user interests during the testing phase. To address these issues, we propose T$^2$ARec, a novel model leveraging state space model for TTT by introducing two Test-Time Alignment modules tailored for sequential recommendation, effectively capturing the distribution shifts in user interest patterns over time. Specifically, T$^2$ARec aligns absolute time intervals with model-adaptive learning intervals to capture temporal dynamics and introduce an interest state alignment mechanism to effectively and explicitly identify the user interest shifts with theoretical guarantees. These two alignment modules enable efficient and incremental updates to model parameters in a self-supervised manner during testing, enhancing predictions for online recommendation. Extensive evaluations on three benchmark datasets demonstrate that T$^2$ARec achieves state-of-the-art performance and robustly mitigates the challenges posed by user interest shifts.

Challenging the Boundaries of Reasoning: An Olympiad-Level Math Benchmark for Large Language Models

In recent years, the rapid development of large reasoning models has resulted in the saturation of existing benchmarks for evaluating mathematical reasoning, highlighting the urgent need for more challenging and rigorous evaluation frameworks. To address this gap, we introduce OlymMATH, a novel Olympiad-level mathematical benchmark, designed to rigorously test the complex reasoning capabilities of LLMs. OlymMATH features 200 meticulously curated problems, each manually verified and available in parallel English and Chinese versions. The problems are systematically organized into two distinct difficulty tiers: (1) AIME-level problems (easy) that establish a baseline for mathematical reasoning assessment, and (2) significantly more challenging problems (hard) designed to push the boundaries of current state-of-the-art models. In our benchmark, these problems span four core mathematical fields, each including a verifiable numerical solution to enable objective, rule-based evaluation. Empirical results underscore the significant challenge presented by OlymMATH, with state-of-the-art models including DeepSeek-R1 and OpenAI's o3-mini demonstrating notably limited accuracy on the hard subset. Furthermore, the benchmark facilitates comprehensive bilingual assessment of mathematical reasoning abilities-a critical dimension that remains largely unaddressed in mainstream mathematical reasoning benchmarks. We release the OlymMATH benchmark at the STILL project: https://github.com/RUCAIBox/Slow_Thinking_with_LLMs.